In PromptMR Plus, the required_grad of the prompt parameter is controlled by the learnable_prompt configuration parameter, which defaults to false. This seems to treat the prompt as an initial noise tensor in DIP.
I tried to train with learnable_prompt: false set unchanged(false) and the results were good (SSIM 0.975+)
Does this mean that the PromptBlock is not used to store general figure features of the dataset?
I'm confused. Can anyone clarify this for me?